- Published on
IPhone 18 Pro Series Officially Goes on Sale, New AI Features 'Self-Contradictory'
- Authors
- Name
- DeepHorus
At 8 a.m. today, Apple's iPhone 18 Pro series officially went on sale, with long lines forming at stores in many locations. Since the highly anticipated iPhone Duo won't arrive until October, the enthusiasm for the iPhone 18 Pro and iPhone 18 Pro Max is relatively subdued; the former's price had already dropped below the launch price on e-commerce platforms during the pre-sale period. Meanwhile, the iOS 27 official version pushed three days ago has sparked controversy over a new feature.
A Reddit user fed a photo of a horse to Apple's new Spatial Reframing feature, intending only to fill out the frame, but a stranger inexplicably appeared in the photo. In the original image, there was only a small hat-like object behind the horse; when adjusting the perspective, the feature generated a nonexistent man based on that small area. TechRadar's testers also found that when processing buildings and busy streets, the AI occasionally adds windows and ledges that don't exist in reality.
Spatial Reframing is a generative photo-editing feature added in iOS 27 that lets users drag the image to simulate the observer's angle after the photographer moves. Because the camera doesn't actually move, the system must generate pixels not covered by the original lens, making "creation out of nothing" inevitable. Although Apple can optimize perspective and lighting, it cannot eliminate the fundamental contradiction: as long as the sensor hasn't recorded those pixels, the model can only guess based on the existing image.
However, Apple is not entirely unprepared. On the iPhone 18 Pro series, Apple introduced the Apple Reference Image feature. This is an optional Reference shooting mode. When enabled, the sensor signs the pixel data directly before firmware modifies the image, generating a hardware-signed DNG "digital negative" that is bound to the main photo processed by conventional computational photography. This proof is established from the moment the phone leaves the factory; the main camera sensor generates an encrypted identity bound to the Secure Enclave.
When the user needs verification, this negative is developed into a tamper-proof reference image via Private Cloud Compute in an auditable environment. Take the Reddit user's horse photo as an example: if shot in normal mode and edited with Spatial Reframing, the photo library contains only the edited final image, with no sensor signature. In Reference mode, however, the same photo project exists on three levels: the edited final image, the photo after undoing the edit, and the Reference Image generated from the sensor recordingâthe latter can prove that the sensor never saw that stranger.
This design seems contradictory: Apple is vigorously promoting generative AI on one hand, while offering hardware-level verification against AI on the other. In fact, Reference Image pushes the starting point of trust to the sensor, so that the "original record" is no longer self-claimed by the photo library or operating system. It re-establishes a boundary with the sensor as the core of truth: within the boundary is pixel data formed by optical signals; outside the boundary may be hallucinated content generated by AI.
In the future, photos may no longer be simply divided into "original/edited"; instead, there will be three states: no Reference Image, Reference Image present but the photo has been edited, and a valid Reference Image consistent with the sensor recording. This dual-track approach may extend to more fieldsâwhen AI can clone voices, recording devices may need hardware-signed original audio tracks; when models can correct wearable device data, medical devices may also need to retain underlying records. Hardware will not only be responsible for capture, but also prove that "the above data came from me."
